Amell et al. (2025) Probabilistic Near‐Real‐Time Retrievals of Rain Over Africa Using Deep Learning

⚠️ Warning: This summary was generated from the abstract only, as the full text was not available.

Identification

• Journal: Journal of Geophysical Research Atmospheres
• Year: 2025
• Date: 2025-10-18
• Authors: Adrià Amell, Lilian Hee, Simon Pfreundschuh, Patrick Eriksson
• DOI: 10.1029/2025jd044595

Research Groups

Not available from the provided abstract.

Short Summary

This paper introduces Rain over Africa (RoA), a public, near-real-time precipitation retrieval algorithm for the African continent based on Meteosat thermal infrared observations. RoA provides precipitation estimates with low latency and detailed uncertainty descriptions, demonstrating accuracy comparable to slower methods and improved timeliness over established products like IMERG for land regions.

Objective

• To introduce and evaluate Rain over Africa (RoA), a new public algorithm for near-real-time precipitation estimation over the African continent, utilizing Meteosat thermal infrared observations and a convolutional and quantile regression neural network.

Study Configuration

• Spatial Scale: African continent and surrounding ocean regions, with an effective resolution of 30 kilometres.
• Temporal Scale: Near-real-time, with precipitation monitoring constantly minutes after input data dissemination; trained and evaluated using four years of data.

Methodology and Data

• Models used: Convolutional and quantile regression neural network.
• Data sources: Meteosat thermal infrared observations (for RoA retrievals); calibration satellite data from the Global Precipitation Measurement (GPM) mission (for training and evaluation labels).

Main Results

• RoA provides near-real-time precipitation estimates with low latency (minutes after data dissemination) and accuracy comparable to estimates requiring hours or more to obtain.
• The algorithm enables detailed, case-specific descriptions of retrieval uncertainty through its probabilistic nature, allowing for the use of probabilities of exceeding precipitation thresholds.
• RoA overcomes limitations observed in earlier near-real-time retrievals for Africa and can run on regular workstations.
• Over land, RoA retrievals offer a 30-kilometre effective resolution, providing more timely and detailed information than established IMERG precipitation estimates.
• RoA is also applicable over surrounding ocean regions, maintaining similar performance, although IMERG exhibits a better effective resolution under its more favorable conditions.
• Assessment reveals similar diurnal cycles between RoA and IMERG, with RoA showing more consistency compared to some instability in IMERG.
• An annual mean analysis, including CHIRPS estimates, indicates regional differences among the three products, with RoA not exhibiting clear outlier behavior.

Contributions

• Introduction of RoA, a novel public, near-real-time, low-latency precipitation retrieval algorithm specifically designed for the African continent.
• Implementation of a convolutional and quantile regression neural network to provide detailed, case-specific descriptions of retrieval uncertainty, addressing the inherent uncertainties of satellite precipitation retrievals.
• Demonstrated capability to overcome limitations of earlier near-real-time retrievals for Africa.
• Provision of more timely and detailed precipitation estimates over land compared to established products like IMERG, with a 30-kilometre effective resolution.
• The algorithm's efficiency allows it to run on regular workstations, enhancing accessibility.

Funding

Not available from the provided abstract.

Citation

@article{Amell2025Probabilistic,
  author = {Amell, Adrià and Hee, Lilian and Pfreundschuh, Simon and Eriksson, Patrick},
  title = {Probabilistic Near‐Real‐Time Retrievals of Rain Over Africa Using Deep Learning},
  journal = {Journal of Geophysical Research Atmospheres},
  year = {2025},
  doi = {10.1029/2025jd044595},
  url = {https://doi.org/10.1029/2025jd044595}
}